1 // SPDX-License-Identifier: GPL-2.0
4 #include "bkey_methods.h"
6 #include "btree_cache.h"
8 #include "btree_iter.h"
9 #include "btree_locking.h"
10 #include "btree_update.h"
11 #include "btree_update_interior.h"
18 #include "journal_reclaim.h"
19 #include "journal_seq_blacklist.h"
22 #include <linux/sched/mm.h>
23 #include <trace/events/bcachefs.h>
25 static void verify_no_dups(struct btree *b,
26 struct bkey_packed *start,
27 struct bkey_packed *end)
29 #ifdef CONFIG_BCACHEFS_DEBUG
30 struct bkey_packed *k, *p;
35 for (p = start, k = bkey_next_skip_noops(start, end);
37 p = k, k = bkey_next_skip_noops(k, end)) {
38 struct bkey l = bkey_unpack_key(b, p);
39 struct bkey r = bkey_unpack_key(b, k);
41 BUG_ON(bkey_cmp(l.p, bkey_start_pos(&r)) >= 0);
46 static void set_needs_whiteout(struct bset *i, int v)
48 struct bkey_packed *k;
52 k = bkey_next_skip_noops(k, vstruct_last(i)))
53 k->needs_whiteout = v;
56 static void btree_bounce_free(struct bch_fs *c, size_t size,
57 bool used_mempool, void *p)
60 mempool_free(p, &c->btree_bounce_pool);
65 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
68 unsigned flags = memalloc_nofs_save();
71 BUG_ON(size > btree_bytes(c));
73 *used_mempool = false;
74 p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
77 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
79 memalloc_nofs_restore(flags);
83 static void sort_bkey_ptrs(const struct btree *bt,
84 struct bkey_packed **ptrs, unsigned nr)
86 unsigned n = nr, a = nr / 2, b, c, d;
91 /* Heap sort: see lib/sort.c: */
96 swap(ptrs[0], ptrs[n]);
100 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
101 b = bch2_bkey_cmp_packed(bt,
103 ptrs[d]) >= 0 ? c : d;
108 bch2_bkey_cmp_packed(bt,
115 swap(ptrs[b], ptrs[c]);
120 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
122 struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
123 bool used_mempool = false;
124 size_t bytes = b->whiteout_u64s * sizeof(u64);
126 if (!b->whiteout_u64s)
129 new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
131 ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
133 for (k = unwritten_whiteouts_start(c, b);
134 k != unwritten_whiteouts_end(c, b);
138 sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
142 while (ptrs != ptrs_end) {
148 verify_no_dups(b, new_whiteouts,
149 (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
151 memcpy_u64s(unwritten_whiteouts_start(c, b),
152 new_whiteouts, b->whiteout_u64s);
154 btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
157 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
158 bool compacting, enum compact_mode mode)
160 if (!bset_dead_u64s(b, t))
165 return should_compact_bset_lazy(b, t) ||
166 (compacting && !bset_written(b, bset(b, t)));
174 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
179 for_each_bset(b, t) {
180 struct bset *i = bset(b, t);
181 struct bkey_packed *k, *n, *out, *start, *end;
182 struct btree_node_entry *src = NULL, *dst = NULL;
184 if (t != b->set && !bset_written(b, i)) {
185 src = container_of(i, struct btree_node_entry, keys);
186 dst = max(write_block(b),
187 (void *) btree_bkey_last(b, t - 1));
193 if (!should_compact_bset(b, t, ret, mode)) {
195 memmove(dst, src, sizeof(*src) +
196 le16_to_cpu(src->keys.u64s) *
199 set_btree_bset(b, t, i);
204 start = btree_bkey_first(b, t);
205 end = btree_bkey_last(b, t);
208 memmove(dst, src, sizeof(*src));
210 set_btree_bset(b, t, i);
215 for (k = start; k != end; k = n) {
216 n = bkey_next_skip_noops(k, end);
218 if (!bkey_deleted(k)) {
220 out = bkey_next(out);
222 BUG_ON(k->needs_whiteout);
226 i->u64s = cpu_to_le16((u64 *) out - i->_data);
227 set_btree_bset_end(b, t);
228 bch2_bset_set_no_aux_tree(b, t);
232 bch2_verify_btree_nr_keys(b);
234 bch2_btree_build_aux_trees(b);
239 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
240 enum compact_mode mode)
242 return bch2_drop_whiteouts(b, mode);
245 static void btree_node_sort(struct bch_fs *c, struct btree *b,
246 struct btree_iter *iter,
249 bool filter_whiteouts)
251 struct btree_node *out;
252 struct sort_iter sort_iter;
254 struct bset *start_bset = bset(b, &b->set[start_idx]);
255 bool used_mempool = false;
256 u64 start_time, seq = 0;
257 unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
258 bool sorting_entire_node = start_idx == 0 &&
261 sort_iter_init(&sort_iter, b);
263 for (t = b->set + start_idx;
264 t < b->set + end_idx;
266 u64s += le16_to_cpu(bset(b, t)->u64s);
267 sort_iter_add(&sort_iter,
268 btree_bkey_first(b, t),
269 btree_bkey_last(b, t));
272 bytes = sorting_entire_node
274 : __vstruct_bytes(struct btree_node, u64s);
276 out = btree_bounce_alloc(c, bytes, &used_mempool);
278 start_time = local_clock();
280 u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
282 out->keys.u64s = cpu_to_le16(u64s);
284 BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
286 if (sorting_entire_node)
287 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
290 /* Make sure we preserve bset journal_seq: */
291 for (t = b->set + start_idx; t < b->set + end_idx; t++)
292 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
293 start_bset->journal_seq = cpu_to_le64(seq);
295 if (sorting_entire_node) {
296 unsigned u64s = le16_to_cpu(out->keys.u64s);
298 BUG_ON(bytes != btree_bytes(c));
301 * Our temporary buffer is the same size as the btree node's
302 * buffer, we can just swap buffers instead of doing a big
306 out->keys.u64s = cpu_to_le16(u64s);
308 set_btree_bset(b, b->set, &b->data->keys);
310 start_bset->u64s = out->keys.u64s;
311 memcpy_u64s(start_bset->start,
313 le16_to_cpu(out->keys.u64s));
316 for (i = start_idx + 1; i < end_idx; i++)
317 b->nr.bset_u64s[start_idx] +=
322 for (i = start_idx + 1; i < b->nsets; i++) {
323 b->nr.bset_u64s[i] = b->nr.bset_u64s[i + shift];
324 b->set[i] = b->set[i + shift];
327 for (i = b->nsets; i < MAX_BSETS; i++)
328 b->nr.bset_u64s[i] = 0;
330 set_btree_bset_end(b, &b->set[start_idx]);
331 bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
333 btree_bounce_free(c, bytes, used_mempool, out);
335 bch2_verify_btree_nr_keys(b);
338 void bch2_btree_sort_into(struct bch_fs *c,
342 struct btree_nr_keys nr;
343 struct btree_node_iter src_iter;
344 u64 start_time = local_clock();
346 BUG_ON(dst->nsets != 1);
348 bch2_bset_set_no_aux_tree(dst, dst->set);
350 bch2_btree_node_iter_init_from_start(&src_iter, src);
352 if (btree_node_is_extents(src))
353 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
358 nr = bch2_sort_repack(btree_bset_first(dst),
363 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
366 set_btree_bset_end(dst, dst->set);
368 dst->nr.live_u64s += nr.live_u64s;
369 dst->nr.bset_u64s[0] += nr.bset_u64s[0];
370 dst->nr.packed_keys += nr.packed_keys;
371 dst->nr.unpacked_keys += nr.unpacked_keys;
373 bch2_verify_btree_nr_keys(dst);
376 #define SORT_CRIT (4096 / sizeof(u64))
379 * We're about to add another bset to the btree node, so if there's currently
380 * too many bsets - sort some of them together:
382 static bool btree_node_compact(struct bch_fs *c, struct btree *b,
383 struct btree_iter *iter)
385 unsigned unwritten_idx;
388 for (unwritten_idx = 0;
389 unwritten_idx < b->nsets;
391 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
394 if (b->nsets - unwritten_idx > 1) {
395 btree_node_sort(c, b, iter, unwritten_idx,
400 if (unwritten_idx > 1) {
401 btree_node_sort(c, b, iter, 0, unwritten_idx, false);
408 void bch2_btree_build_aux_trees(struct btree *b)
413 bch2_bset_build_aux_tree(b, t,
414 !bset_written(b, bset(b, t)) &&
415 t == bset_tree_last(b));
419 * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
422 * Safe to call if there already is an unwritten bset - will only add a new bset
423 * if @b doesn't already have one.
425 * Returns true if we sorted (i.e. invalidated iterators
427 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
428 struct btree_iter *iter)
430 struct btree_node_entry *bne;
433 EBUG_ON(!(b->c.lock.state.seq & 1));
434 EBUG_ON(iter && iter->l[b->c.level].b != b);
436 did_sort = btree_node_compact(c, b, iter);
438 bne = want_new_bset(c, b);
440 bch2_bset_init_next(c, b, bne);
442 bch2_btree_build_aux_trees(b);
444 if (iter && did_sort)
445 bch2_btree_iter_reinit_node(iter, b);
448 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
451 pr_buf(out, "%s level %u/%u\n ",
452 bch2_btree_ids[b->c.btree_id],
454 c->btree_roots[b->c.btree_id].level);
455 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
458 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
460 struct btree *b, struct bset *i,
461 unsigned offset, int write)
463 pr_buf(out, "error validating btree node ");
465 pr_buf(out, "before write ");
467 pr_buf(out, "on %s ", ca->name);
468 pr_buf(out, "at btree ");
469 btree_pos_to_text(out, c, b);
471 pr_buf(out, "\n node offset %u", b->written);
473 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
476 enum btree_err_type {
478 BTREE_ERR_WANT_RETRY,
479 BTREE_ERR_MUST_RETRY,
483 enum btree_validate_ret {
484 BTREE_RETRY_READ = 64,
487 #define btree_err(type, c, ca, b, i, msg, ...) \
491 char *_buf2 = _buf; \
492 struct printbuf out = PBUF(_buf); \
494 _buf2 = kmalloc(4096, GFP_ATOMIC); \
496 out = _PBUF(_buf2, 4986); \
498 btree_err_msg(&out, c, ca, b, i, b->written, write); \
499 pr_buf(&out, ": " msg, ##__VA_ARGS__); \
501 if (type == BTREE_ERR_FIXABLE && \
503 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) { \
504 mustfix_fsck_err(c, "%s", _buf2); \
510 bch_err(c, "%s", _buf2); \
513 case BTREE_ERR_FIXABLE: \
514 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
516 case BTREE_ERR_WANT_RETRY: \
518 ret = BTREE_RETRY_READ; \
522 case BTREE_ERR_MUST_RETRY: \
523 ret = BTREE_RETRY_READ; \
525 case BTREE_ERR_FATAL: \
526 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
531 bch_err(c, "corrupt metadata before write: %s", _buf2); \
533 if (bch2_fs_inconsistent(c)) { \
534 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
545 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
547 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
548 struct btree *b, struct bset *i,
549 unsigned sectors, int write, bool have_retry)
551 unsigned version = le16_to_cpu(i->version);
557 btree_err_on((version != BCH_BSET_VERSION_OLD &&
558 version < bcachefs_metadata_version_min) ||
559 version >= bcachefs_metadata_version_max,
560 BTREE_ERR_FATAL, c, ca, b, i,
561 "unsupported bset version");
563 if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
564 BTREE_ERR_FIXABLE, c, ca, b, i,
565 "bset past end of btree node")) {
570 btree_err_on(b->written && !i->u64s,
571 BTREE_ERR_FIXABLE, c, ca, b, i,
575 struct btree_node *bn =
576 container_of(i, struct btree_node, keys);
577 /* These indicate that we read the wrong btree node: */
579 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
580 struct bch_btree_ptr_v2 *bp =
581 &bkey_i_to_btree_ptr_v2(&b->key)->v;
584 btree_err_on(bp->seq != bn->keys.seq,
585 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
586 "incorrect sequence number (wrong btree node)");
589 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
590 BTREE_ERR_MUST_RETRY, c, ca, b, i,
591 "incorrect btree id");
593 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
594 BTREE_ERR_MUST_RETRY, c, ca, b, i,
597 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
598 u64 *p = (u64 *) &bn->ptr;
604 compat_btree_node(b->c.level, b->c.btree_id, version,
605 BSET_BIG_ENDIAN(i), write, bn);
607 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
608 struct bch_btree_ptr_v2 *bp =
609 &bkey_i_to_btree_ptr_v2(&b->key)->v;
611 if (BTREE_PTR_RANGE_UPDATED(bp)) {
612 b->data->min_key = bp->min_key;
613 b->data->max_key = b->key.k.p;
616 btree_err_on(bkey_cmp(b->data->min_key, bp->min_key),
617 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
618 "incorrect min_key: got %s should be %s",
619 (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
620 (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
623 btree_err_on(bkey_cmp(bn->max_key, b->key.k.p),
624 BTREE_ERR_MUST_RETRY, c, ca, b, i,
625 "incorrect max key %s",
626 (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
629 compat_btree_node(b->c.level, b->c.btree_id, version,
630 BSET_BIG_ENDIAN(i), write, bn);
632 err = bch2_bkey_format_validate(&bn->format);
634 BTREE_ERR_FATAL, c, ca, b, i,
635 "invalid bkey format: %s", err);
637 compat_bformat(b->c.level, b->c.btree_id, version,
638 BSET_BIG_ENDIAN(i), write,
645 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
646 struct bset *i, unsigned *whiteout_u64s,
647 int write, bool have_retry)
649 unsigned version = le16_to_cpu(i->version);
650 struct bkey_packed *k, *prev = NULL;
651 bool seen_non_whiteout = false;
654 if (!BSET_SEPARATE_WHITEOUTS(i)) {
655 seen_non_whiteout = true;
660 k != vstruct_last(i);) {
665 if (btree_err_on(bkey_next(k) > vstruct_last(i),
666 BTREE_ERR_FIXABLE, c, NULL, b, i,
667 "key extends past end of bset")) {
668 i->u64s = cpu_to_le16((u64 *) k - i->_data);
672 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
673 BTREE_ERR_FIXABLE, c, NULL, b, i,
674 "invalid bkey format %u", k->format)) {
675 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
676 memmove_u64s_down(k, bkey_next(k),
677 (u64 *) vstruct_end(i) - (u64 *) k);
681 /* XXX: validate k->u64s */
683 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
684 BSET_BIG_ENDIAN(i), write,
687 u = __bkey_disassemble(b, k, &tmp);
689 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
690 bch2_bkey_in_btree_node(b, u.s_c) ?:
691 (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
695 bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
696 btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
697 "invalid bkey: %s\n%s", invalid, buf);
699 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
700 memmove_u64s_down(k, bkey_next(k),
701 (u64 *) vstruct_end(i) - (u64 *) k);
706 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
707 BSET_BIG_ENDIAN(i), write,
711 * with the separate whiteouts thing (used for extents), the
712 * second set of keys actually can have whiteouts too, so we
713 * can't solely go off bkey_deleted()...
716 if (!seen_non_whiteout &&
718 (prev && bkey_iter_cmp(b, prev, k) > 0))) {
719 *whiteout_u64s = k->_data - i->_data;
720 seen_non_whiteout = true;
721 } else if (prev && bkey_iter_cmp(b, prev, k) > 0) {
724 struct bkey up = bkey_unpack_key(b, prev);
726 bch2_bkey_to_text(&PBUF(buf1), &up);
727 bch2_bkey_to_text(&PBUF(buf2), u.k);
729 bch2_dump_bset(c, b, i, 0);
730 btree_err(BTREE_ERR_FATAL, c, NULL, b, i,
731 "keys out of order: %s > %s",
733 /* XXX: repair this */
737 k = bkey_next_skip_noops(k, vstruct_last(i));
743 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
744 struct btree *b, bool have_retry)
746 struct btree_node_entry *bne;
747 struct sort_iter *iter;
748 struct btree_node *sorted;
749 struct bkey_packed *k;
750 struct bch_extent_ptr *ptr;
752 bool used_mempool, blacklisted;
754 int ret, retry_read = 0, write = READ;
756 iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
757 sort_iter_init(iter, b);
758 iter->size = (btree_blocks(c) + 1) * 2;
760 if (bch2_meta_read_fault("btree"))
761 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
764 btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
765 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
768 btree_err_on(!b->data->keys.seq,
769 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
772 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
773 struct bch_btree_ptr_v2 *bp =
774 &bkey_i_to_btree_ptr_v2(&b->key)->v;
776 btree_err_on(b->data->keys.seq != bp->seq,
777 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
778 "got wrong btree node (seq %llx want %llx)",
779 b->data->keys.seq, bp->seq);
782 while (b->written < c->opts.btree_node_size) {
783 unsigned sectors, whiteout_u64s = 0;
785 struct bch_csum csum;
786 bool first = !b->written;
791 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
792 BTREE_ERR_WANT_RETRY, c, ca, b, i,
793 "unknown checksum type %llu",
796 nonce = btree_nonce(i, b->written << 9);
797 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
799 btree_err_on(bch2_crc_cmp(csum, b->data->csum),
800 BTREE_ERR_WANT_RETRY, c, ca, b, i,
803 bset_encrypt(c, i, b->written << 9);
805 btree_err_on(btree_node_is_extents(b) &&
806 !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
807 BTREE_ERR_FATAL, c, NULL, b, NULL,
808 "btree node does not have NEW_EXTENT_OVERWRITE set");
810 sectors = vstruct_sectors(b->data, c->block_bits);
812 bne = write_block(b);
815 if (i->seq != b->data->keys.seq)
818 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
819 BTREE_ERR_WANT_RETRY, c, ca, b, i,
820 "unknown checksum type %llu",
823 nonce = btree_nonce(i, b->written << 9);
824 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
826 btree_err_on(bch2_crc_cmp(csum, bne->csum),
827 BTREE_ERR_WANT_RETRY, c, ca, b, i,
830 bset_encrypt(c, i, b->written << 9);
832 sectors = vstruct_sectors(bne, c->block_bits);
835 ret = validate_bset(c, ca, b, i, sectors,
841 btree_node_set_format(b, b->data->format);
843 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
848 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
850 b->written += sectors;
852 blacklisted = bch2_journal_seq_is_blacklisted(c,
853 le64_to_cpu(i->journal_seq),
856 btree_err_on(blacklisted && first,
857 BTREE_ERR_FIXABLE, c, ca, b, i,
858 "first btree node bset has blacklisted journal seq");
859 if (blacklisted && !first)
862 sort_iter_add(iter, i->start,
863 vstruct_idx(i, whiteout_u64s));
866 vstruct_idx(i, whiteout_u64s),
870 for (bne = write_block(b);
871 bset_byte_offset(b, bne) < btree_bytes(c);
872 bne = (void *) bne + block_bytes(c))
873 btree_err_on(bne->keys.seq == b->data->keys.seq,
874 BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
875 "found bset signature after last bset");
877 sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
878 sorted->keys.u64s = 0;
880 set_btree_bset(b, b->set, &b->data->keys);
882 b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
884 u64s = le16_to_cpu(sorted->keys.u64s);
886 sorted->keys.u64s = cpu_to_le16(u64s);
887 swap(sorted, b->data);
888 set_btree_bset(b, b->set, &b->data->keys);
891 BUG_ON(b->nr.live_u64s != u64s);
893 btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
896 for (k = i->start; k != vstruct_last(i);) {
898 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
899 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
902 (bch2_inject_invalid_keys &&
903 !bversion_cmp(u.k->version, MAX_VERSION))) {
906 bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
907 btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
908 "invalid bkey %s: %s", buf, invalid);
910 btree_keys_account_key_drop(&b->nr, 0, k);
912 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
913 memmove_u64s_down(k, bkey_next(k),
914 (u64 *) vstruct_end(i) - (u64 *) k);
915 set_btree_bset_end(b, b->set);
919 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
920 struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
925 k = bkey_next_skip_noops(k, vstruct_last(i));
928 bch2_bset_build_aux_tree(b, b->set, false);
930 set_needs_whiteout(btree_bset_first(b), true);
932 btree_node_reset_sib_u64s(b);
934 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
935 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
937 if (ca->mi.state != BCH_MEMBER_STATE_rw)
938 set_btree_node_need_rewrite(b);
941 mempool_free(iter, &c->fill_iter);
944 if (ret == BTREE_RETRY_READ) {
947 bch2_inconsistent_error(c);
948 set_btree_node_read_error(b);
953 static void btree_node_read_work(struct work_struct *work)
955 struct btree_read_bio *rb =
956 container_of(work, struct btree_read_bio, work);
957 struct bch_fs *c = rb->c;
958 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
959 struct btree *b = rb->bio.bi_private;
960 struct bio *bio = &rb->bio;
961 struct bch_io_failures failed = { .nr = 0 };
968 bch_info(c, "retrying read");
969 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
970 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
972 bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
973 bio->bi_iter.bi_sector = rb->pick.ptr.offset;
974 bio->bi_iter.bi_size = btree_bytes(c);
976 if (rb->have_ioref) {
977 bio_set_dev(bio, ca->disk_sb.bdev);
978 submit_bio_wait(bio);
980 bio->bi_status = BLK_STS_REMOVED;
984 btree_pos_to_text(&out, c, b);
985 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
986 bch2_blk_status_to_str(bio->bi_status), buf);
988 percpu_ref_put(&ca->io_ref);
989 rb->have_ioref = false;
991 bch2_mark_io_failure(&failed, &rb->pick);
993 can_retry = bch2_bkey_pick_read_device(c,
994 bkey_i_to_s_c(&b->key),
995 &failed, &rb->pick) > 0;
997 if (!bio->bi_status &&
998 !bch2_btree_node_read_done(c, ca, b, can_retry))
1002 set_btree_node_read_error(b);
1007 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1010 clear_btree_node_read_in_flight(b);
1011 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1014 static void btree_node_read_endio(struct bio *bio)
1016 struct btree_read_bio *rb =
1017 container_of(bio, struct btree_read_bio, bio);
1018 struct bch_fs *c = rb->c;
1020 if (rb->have_ioref) {
1021 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1022 bch2_latency_acct(ca, rb->start_time, READ);
1025 queue_work(system_unbound_wq, &rb->work);
1028 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1031 struct extent_ptr_decoded pick;
1032 struct btree_read_bio *rb;
1037 trace_btree_read(c, b);
1039 ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1041 if (bch2_fs_fatal_err_on(ret <= 0, c,
1042 "btree node read error: no device to read from")) {
1043 set_btree_node_read_error(b);
1047 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1049 bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1052 rb = container_of(bio, struct btree_read_bio, bio);
1054 rb->start_time = local_clock();
1055 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1057 INIT_WORK(&rb->work, btree_node_read_work);
1058 bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
1059 bio->bi_iter.bi_sector = pick.ptr.offset;
1060 bio->bi_end_io = btree_node_read_endio;
1061 bio->bi_private = b;
1062 bch2_bio_map(bio, b->data, btree_bytes(c));
1064 set_btree_node_read_in_flight(b);
1066 if (rb->have_ioref) {
1067 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1069 bio_set_dev(bio, ca->disk_sb.bdev);
1072 submit_bio_wait(bio);
1074 bio->bi_private = b;
1075 btree_node_read_work(&rb->work);
1080 bio->bi_status = BLK_STS_REMOVED;
1083 btree_node_read_work(&rb->work);
1085 queue_work(system_unbound_wq, &rb->work);
1090 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1091 const struct bkey_i *k, unsigned level)
1097 closure_init_stack(&cl);
1100 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1104 b = bch2_btree_node_mem_alloc(c);
1105 bch2_btree_cache_cannibalize_unlock(c);
1109 bkey_copy(&b->key, k);
1110 BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1112 bch2_btree_node_read(c, b, true);
1114 if (btree_node_read_error(b)) {
1115 bch2_btree_node_hash_remove(&c->btree_cache, b);
1117 mutex_lock(&c->btree_cache.lock);
1118 list_move(&b->list, &c->btree_cache.freeable);
1119 mutex_unlock(&c->btree_cache.lock);
1125 bch2_btree_set_root_for_read(c, b);
1127 six_unlock_write(&b->c.lock);
1128 six_unlock_intent(&b->c.lock);
1133 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1134 struct btree_write *w)
1136 unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1144 } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1147 closure_put(&((struct btree_update *) new)->cl);
1149 bch2_journal_pin_drop(&c->journal, &w->journal);
1152 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1154 struct btree_write *w = btree_prev_write(b);
1156 bch2_btree_complete_write(c, b, w);
1157 btree_node_io_unlock(b);
1160 static void bch2_btree_node_write_error(struct bch_fs *c,
1161 struct btree_write_bio *wbio)
1163 struct btree *b = wbio->wbio.bio.bi_private;
1165 struct bch_extent_ptr *ptr;
1166 struct btree_trans trans;
1167 struct btree_iter *iter;
1170 bch2_bkey_buf_init(&k);
1171 bch2_trans_init(&trans, c, 0, 0);
1173 iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1174 BTREE_MAX_DEPTH, b->c.level, 0);
1176 ret = bch2_btree_iter_traverse(iter);
1180 /* has node been freed? */
1181 if (iter->l[b->c.level].b != b) {
1182 /* node has been freed: */
1183 BUG_ON(!btree_node_dying(b));
1187 BUG_ON(!btree_node_hashed(b));
1189 bch2_bkey_buf_copy(&k, c, &b->key);
1191 bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1192 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1194 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1197 ret = bch2_btree_node_update_key(c, iter, b, k.k);
1203 bch2_trans_exit(&trans);
1204 bch2_bkey_buf_exit(&k, c);
1205 bio_put(&wbio->wbio.bio);
1206 btree_node_write_done(c, b);
1209 set_btree_node_noevict(b);
1210 bch2_fs_fatal_error(c, "fatal error writing btree node");
1214 void bch2_btree_write_error_work(struct work_struct *work)
1216 struct bch_fs *c = container_of(work, struct bch_fs,
1217 btree_write_error_work);
1221 spin_lock_irq(&c->btree_write_error_lock);
1222 bio = bio_list_pop(&c->btree_write_error_list);
1223 spin_unlock_irq(&c->btree_write_error_lock);
1228 bch2_btree_node_write_error(c,
1229 container_of(bio, struct btree_write_bio, wbio.bio));
1233 static void btree_node_write_work(struct work_struct *work)
1235 struct btree_write_bio *wbio =
1236 container_of(work, struct btree_write_bio, work);
1237 struct bch_fs *c = wbio->wbio.c;
1238 struct btree *b = wbio->wbio.bio.bi_private;
1240 btree_bounce_free(c,
1242 wbio->wbio.used_mempool,
1245 if (wbio->wbio.failed.nr) {
1246 unsigned long flags;
1248 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1249 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1250 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1252 queue_work(c->wq, &c->btree_write_error_work);
1256 bio_put(&wbio->wbio.bio);
1257 btree_node_write_done(c, b);
1260 static void btree_node_write_endio(struct bio *bio)
1262 struct bch_write_bio *wbio = to_wbio(bio);
1263 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
1264 struct bch_write_bio *orig = parent ?: wbio;
1265 struct bch_fs *c = wbio->c;
1266 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
1267 unsigned long flags;
1269 if (wbio->have_ioref)
1270 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1272 if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1273 bch2_blk_status_to_str(bio->bi_status)) ||
1274 bch2_meta_write_fault("btree")) {
1275 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1276 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1277 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1280 if (wbio->have_ioref)
1281 percpu_ref_put(&ca->io_ref);
1285 bio_endio(&parent->bio);
1287 struct btree_write_bio *wb =
1288 container_of(orig, struct btree_write_bio, wbio);
1290 INIT_WORK(&wb->work, btree_node_write_work);
1291 queue_work(system_unbound_wq, &wb->work);
1295 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1296 struct bset *i, unsigned sectors)
1298 unsigned whiteout_u64s = 0;
1301 if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1304 ret = validate_bset(c, NULL, b, i, sectors, WRITE, false) ?:
1305 validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false);
1307 bch2_inconsistent_error(c);
1314 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1315 enum six_lock_type lock_type_held)
1317 struct btree_write_bio *wbio;
1318 struct bset_tree *t;
1320 struct btree_node *bn = NULL;
1321 struct btree_node_entry *bne = NULL;
1323 struct bch_extent_ptr *ptr;
1324 struct sort_iter sort_iter;
1326 unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1329 unsigned long old, new;
1330 bool validate_before_checksum = false;
1333 bch2_bkey_buf_init(&k);
1335 if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1339 * We may only have a read lock on the btree node - the dirty bit is our
1340 * "lock" against racing with other threads that may be trying to start
1341 * a write, we do a write iff we clear the dirty bit. Since setting the
1342 * dirty bit requires a write lock, we can't race with other threads
1346 old = new = READ_ONCE(b->flags);
1348 if (!(old & (1 << BTREE_NODE_dirty)))
1351 if (!btree_node_may_write(b))
1354 if (old & (1 << BTREE_NODE_never_write))
1357 if (old & (1 << BTREE_NODE_write_in_flight)) {
1358 btree_node_wait_on_io(b);
1362 new &= ~(1 << BTREE_NODE_dirty);
1363 new &= ~(1 << BTREE_NODE_need_write);
1364 new |= (1 << BTREE_NODE_write_in_flight);
1365 new |= (1 << BTREE_NODE_just_written);
1366 new ^= (1 << BTREE_NODE_write_idx);
1367 } while (cmpxchg_acquire(&b->flags, old, new) != old);
1369 atomic_dec(&c->btree_cache.dirty);
1371 BUG_ON(btree_node_fake(b));
1372 BUG_ON((b->will_make_reachable != 0) != !b->written);
1374 BUG_ON(b->written >= c->opts.btree_node_size);
1375 BUG_ON(b->written & (c->opts.block_size - 1));
1376 BUG_ON(bset_written(b, btree_bset_last(b)));
1377 BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1378 BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1380 bch2_sort_whiteouts(c, b);
1382 sort_iter_init(&sort_iter, b);
1385 ? sizeof(struct btree_node)
1386 : sizeof(struct btree_node_entry);
1388 bytes += b->whiteout_u64s * sizeof(u64);
1390 for_each_bset(b, t) {
1393 if (bset_written(b, i))
1396 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1397 sort_iter_add(&sort_iter,
1398 btree_bkey_first(b, t),
1399 btree_bkey_last(b, t));
1400 seq = max(seq, le64_to_cpu(i->journal_seq));
1403 BUG_ON(b->written && !seq);
1405 /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1408 data = btree_bounce_alloc(c, bytes, &used_mempool);
1416 bne->keys = b->data->keys;
1420 i->journal_seq = cpu_to_le64(seq);
1423 sort_iter_add(&sort_iter,
1424 unwritten_whiteouts_start(c, b),
1425 unwritten_whiteouts_end(c, b));
1426 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1428 b->whiteout_u64s = 0;
1430 u64s = bch2_sort_keys(i->start, &sort_iter, false);
1431 le16_add_cpu(&i->u64s, u64s);
1433 set_needs_whiteout(i, false);
1435 /* do we have data to write? */
1436 if (b->written && !i->u64s)
1439 bytes_to_write = vstruct_end(i) - data;
1440 sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1442 memset(data + bytes_to_write, 0,
1443 (sectors_to_write << 9) - bytes_to_write);
1445 BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1446 BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1447 BUG_ON(i->seq != b->data->keys.seq);
1449 i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1450 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1451 : cpu_to_le16(c->sb.version);
1452 SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1454 if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1455 validate_before_checksum = true;
1457 /* validate_bset will be modifying: */
1458 if (le16_to_cpu(i->version) <= bcachefs_metadata_version_inode_btree_change)
1459 validate_before_checksum = true;
1461 /* if we're going to be encrypting, check metadata validity first: */
1462 if (validate_before_checksum &&
1463 validate_bset_for_write(c, b, i, sectors_to_write))
1466 bset_encrypt(c, i, b->written << 9);
1468 nonce = btree_nonce(i, b->written << 9);
1471 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1473 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1475 /* if we're not encrypting, check metadata after checksumming: */
1476 if (!validate_before_checksum &&
1477 validate_bset_for_write(c, b, i, sectors_to_write))
1481 * We handle btree write errors by immediately halting the journal -
1482 * after we've done that, we can't issue any subsequent btree writes
1483 * because they might have pointers to new nodes that failed to write.
1485 * Furthermore, there's no point in doing any more btree writes because
1486 * with the journal stopped, we're never going to update the journal to
1487 * reflect that those writes were done and the data flushed from the
1490 * Also on journal error, the pending write may have updates that were
1491 * never journalled (interior nodes, see btree_update_nodes_written()) -
1492 * it's critical that we don't do the write in that case otherwise we
1493 * will have updates visible that weren't in the journal:
1495 * Make sure to update b->written so bch2_btree_init_next() doesn't
1498 if (bch2_journal_error(&c->journal) ||
1502 trace_btree_write(b, bytes_to_write, sectors_to_write);
1504 wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1505 buf_pages(data, sectors_to_write << 9),
1507 struct btree_write_bio, wbio.bio);
1508 wbio_init(&wbio->wbio.bio);
1510 wbio->bytes = bytes;
1511 wbio->wbio.used_mempool = used_mempool;
1512 wbio->wbio.bio.bi_opf = REQ_OP_WRITE|REQ_META;
1513 wbio->wbio.bio.bi_end_io = btree_node_write_endio;
1514 wbio->wbio.bio.bi_private = b;
1516 bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1519 * If we're appending to a leaf node, we don't technically need FUA -
1520 * this write just needs to be persisted before the next journal write,
1521 * which will be marked FLUSH|FUA.
1523 * Similarly if we're writing a new btree root - the pointer is going to
1524 * be in the next journal entry.
1526 * But if we're writing a new btree node (that isn't a root) or
1527 * appending to a non leaf btree node, we need either FUA or a flush
1528 * when we write the parent with the new pointer. FUA is cheaper than a
1529 * flush, and writes appending to leaf nodes aren't blocking anything so
1530 * just make all btree node writes FUA to keep things sane.
1533 bch2_bkey_buf_copy(&k, c, &b->key);
1535 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(k.k)), ptr)
1536 ptr->offset += b->written;
1538 b->written += sectors_to_write;
1540 /* XXX: submitting IO with btree locks held: */
1541 bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_btree, k.k);
1542 bch2_bkey_buf_exit(&k, c);
1545 set_btree_node_noevict(b);
1546 b->written += sectors_to_write;
1548 btree_bounce_free(c, bytes, used_mempool, data);
1549 btree_node_write_done(c, b);
1553 * Work that must be done with write lock held:
1555 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1557 bool invalidated_iter = false;
1558 struct btree_node_entry *bne;
1559 struct bset_tree *t;
1561 if (!btree_node_just_written(b))
1564 BUG_ON(b->whiteout_u64s);
1566 clear_btree_node_just_written(b);
1569 * Note: immediately after write, bset_written() doesn't work - the
1570 * amount of data we had to write after compaction might have been
1571 * smaller than the offset of the last bset.
1573 * However, we know that all bsets have been written here, as long as
1574 * we're still holding the write lock:
1578 * XXX: decide if we really want to unconditionally sort down to a
1582 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1583 invalidated_iter = true;
1585 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1589 set_needs_whiteout(bset(b, t), true);
1591 bch2_btree_verify(c, b);
1594 * If later we don't unconditionally sort down to a single bset, we have
1595 * to ensure this is still true:
1597 BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1599 bne = want_new_bset(c, b);
1601 bch2_bset_init_next(c, b, bne);
1603 bch2_btree_build_aux_trees(b);
1605 return invalidated_iter;
1609 * Use this one if the node is intent locked:
1611 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1612 enum six_lock_type lock_type_held)
1614 BUG_ON(lock_type_held == SIX_LOCK_write);
1616 if (lock_type_held == SIX_LOCK_intent ||
1617 six_lock_tryupgrade(&b->c.lock)) {
1618 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1620 /* don't cycle lock unnecessarily: */
1621 if (btree_node_just_written(b) &&
1622 six_trylock_write(&b->c.lock)) {
1623 bch2_btree_post_write_cleanup(c, b);
1624 six_unlock_write(&b->c.lock);
1627 if (lock_type_held == SIX_LOCK_read)
1628 six_lock_downgrade(&b->c.lock);
1630 __bch2_btree_node_write(c, b, SIX_LOCK_read);
1634 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
1636 struct bucket_table *tbl;
1637 struct rhash_head *pos;
1642 for_each_cached_btree(b, c, tbl, i, pos)
1643 if (test_bit(flag, &b->flags)) {
1645 wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
1652 void bch2_btree_flush_all_reads(struct bch_fs *c)
1654 __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
1657 void bch2_btree_flush_all_writes(struct bch_fs *c)
1659 __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
1662 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
1664 struct bucket_table *tbl;
1665 struct rhash_head *pos;
1670 for_each_cached_btree(b, c, tbl, i, pos) {
1671 unsigned long flags = READ_ONCE(b->flags);
1673 if (!(flags & (1 << BTREE_NODE_dirty)))
1676 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
1678 (flags & (1 << BTREE_NODE_dirty)) != 0,
1679 (flags & (1 << BTREE_NODE_need_write)) != 0,
1682 !list_empty_careful(&b->write_blocked),
1683 b->will_make_reachable != 0,
1684 b->will_make_reachable & 1);